Intelligent zoning design of concrete-faced rockfill dams using image-parameter fusion enhanced generative adversarial networks

The design of concrete-face rockfill dams (CFRDs) is gradually evolving from digitization to intelligent design, primarily driven by advanced technologies such as generative artificial intelligence (AI). Generative AI offers powerful capabilities for extracting and mining data features and generating new design solutions efficiently through inference. These strengths provide critical technical support for intelligent CFRD design, facilitating the full utilization of design data, and significantly improving design efficiency and quality. However, CFRD design is a highly specialized and complex task, and conventional generative AI techniques often fail to produce professional-grade designs. In response to this challenge, this study explored generative intelligent design methods specifically for the critical task of CFRD profile design. An intelligent design method based on feature-fusion generative adversarial networks (GANs) for CFRD profile solutions is proposed. This approach enables the dense representation and augmentation of design data, GAN model training, and automated evaluation, thereby addressing the key challenge of fusing small-sample multimodal image-parameter data. The effectiveness of the intelligent design method for CFRD profiles was validated through algorithm analysis and case studies. The design efficiency was nearly 10 times higher than that of traditional engineer-driven designs, reducing the time required from 1–2 h to approximately 6 min. The proposed intelligent design approach has great potential and provides valuable insights for the further development of intelligent design of rockfill dams. • Generative intelligent design methods specifically for the critical task of CFRD zoning design is explored. • Image and parameter feature-fusion generative adversarial network is proposed for CFRD zoning design. • The design efficiency of the proposed method was nearly 10 times higher than that of traditional engineer-driven designs. • A two-stage training method for GANs, combining parametric and real data, was developed.